digest: 2026-05-07 (alt — parallel run)

[JustinPerea]

Heads-up: parallel run

PR #3 already covers 2026-05-07 from another concurrent run. I noticed it only after committing and chose not to force-push over claude/digest-2026-05-07 — pushed this work to claude/digest-2026-05-07-alt instead so nothing gets clobbered. Pick whichever digest you want to merge (or cherry-pick across); they touch the same _meta/* files so they will conflict on merge.

Summary

Daily AI research digest for 2026-05-07. 7 verified items, 2 conflicts flagged.

Focus areas covered

• Context engineering (2 items): LongSeeker / Context-ReAct (Qwen3-30B-A3B fine-tune, BrowseComp 61.5% vs. 43.2% best baseline) and LCM / Volt (claims to outperform Claude Code on OOLONG at every length 32K–1M with both running Opus 4.6).
• Agent orchestration (1 item): Uno-Orchestra (RL-learned routing, 77.0% macro pass@1 across 13 benches, +16 pp over strongest workflow at ~10× lower cost).
• Coding agents (2 items): Executable world models for ARC-AGI-3 (refactor toward simpler abstractions, 7/25 fully solved); AuditRepairBench (evaluator-channel leakage explains 55–74% of agent-repair leaderboard rank instability).
• Tools & connectors / safety (1 item): AgentTrust (AGPL-3.0 MCP server; allow/warn/block/review verdicts on tool calls; 95% / 96.7% verdict accuracy; low-ms latency).
• Evaluation / agentic retrieval (1 item): BRIGHT-Pro + RTriever-4B (LoRA fine-tune of Qwen3-Embedding-4B with aspect-aware agentic eval).

Conflicts flagged

1. LCM / Volt (arxiv:2605.04050) vs. Claude Code harness — direct claim that Volt + LCM beats Claude Code on OOLONG at every length 32K–1M with both running Opus 4.6. Logged as compare; awaiting third-party reproduction.
2. AuditRepairBench (arxiv:2605.04624) extends EvalEval (2026-05-04 ledger entry) — sharpens the "agent benchmarks are noisy" finding by identifying evaluator-channel leakage as a specific mechanism (55–74% vs. 7% / 13% controls).

Local-model corner

LongSeeker's base is Qwen3-30B-A3B MoE (30B total / ~3B active). At Q4 GGUF that's ~17–18 GB, leaving ~45 GB free on the M5 Max for KV cache; latency closer to a 3B than a 30B. RTriever-4B is also trivially M5-fit if released.

Sources status

• Surveyed: arXiv cs.AI, HF Daily Papers, HF Blog, Anthropic News + Research, Mistral News, DeepMind Blog, Google Research Blog, Meta AI Blog, AI2 Blog, NVIDIA Developer Blog
• Could not access (retry next run): OpenAI Research (HTTP 403, consistent with 2026-05-04)
• No new in-scope posts: Anthropic Research, Mistral, DeepMind, Google Research, Meta AI

Files

• inbox/2026-05-07.md — digest
• _meta/processed.json — 7 new IDs prepended
• _meta/claims-ledger.md — 7 new tracked claims appended

How this differs from PR #3

• PR digest: 2026-05-07 #3 has 8 items; this PR has 7. Overlap on LongSeeker, Uno-Orchestra, AgentTrust. Unique to this PR: LCM / Volt (direct Claude Code comparison), Executable World Models for ARC-AGI-3, AuditRepairBench, BRIGHT-Pro / RTriever-4B. Unique to PR digest: 2026-05-07 #3: Impossibility Triangle, When Context Hurts, Design Conductor 2.0, vLLM V0→V1 retro, Anthropic Finance Agents (Opus 4.7 / Vals AI 64.37%).

---

Generated by Claude Code